Mistral Medium 3.5

Summary

Mistral Medium 3.5 is Mistral AI's frontier-class mid-tier multimodal model, released April 29, 2026 as open weights under a Modified MIT license. It scores 77.6% on SWE-Bench Verified — ahead of Devstral 2 and Qwen3.5 397B A17B — at aggressive $1.50/$7.50 per million-token pricing.

Overview

Mistral Medium 3.5 (mistral-medium-3-5) is Mistral AI's frontier-class mid-tier multimodal model, released April 29, 2026. It is positioned as Mistral's primary play for agentic and coding workloads, with adjustable reasoning depth via a reasoning_effort parameter that lets developers trade latency and cost against capability on the same model. Medium 3.5 is released as open weights under a Modified MIT license — preserving Mistral's open-source ecosystem positioning while adding terms specific to certain commercial deployments.

The headline result is performance: Medium 3.5 scores 77.6% on SWE-Bench Verified, ahead of Devstral 2 and Alibaba's Qwen3.5 397B A17B model. At $1.50/$7.50 per million input/output tokens, Medium 3.5 is materially cheaper than the closed-frontier models (GPT-5.5, Opus 4.7) it competes with on coding-task quality, making it the most aggressive open-weight pricing point at this capability tier.

Specifications

  • Developer: Mistral AI
  • Model String: mistral-medium-3-5
  • Release Date: April 29, 2026
  • Type: Frontier-class multimodal mid-tier LLM with adjustable reasoning
  • License: Modified MIT (open weights, with some commercial-use terms)
  • Reasoning Control: reasoning_effort parameter — trade latency/cost vs. capability
  • Pricing: $1.50 per million input tokens / $7.50 per million output tokens
  • Access: Mistral API; open-weight downloads (Hugging Face and Mistral platform)

Capabilities

Coding (headline result): 77.6% on SWE-Bench Verified — ahead of Devstral 2 and Qwen3.5 397B A17B. Positions Medium 3.5 as the highest-quality open-weight coding model in its release window.

Adjustable Reasoning: reasoning_effort parameter lets developers choose between low-latency answers and deeper deliberation on the same model — analogous to OpenAI's reasoning_effort and Anthropic's thinking modes.

Multimodal: Frontier-class multimodal capability — text and image input.

Agentic Workflows: Mistral's positioning emphasizes agentic and coding use cases as the primary deployment target — with the SWE-Bench result as the headline validation.

Open-Weight Distribution: Modified MIT license allows commercial deployment with some terms — more permissive than Llama's custom license but less permissive than the pure-MIT approach DeepSeek uses for V4.

Mid-Tier Pricing: $1.50/$7.50 per million input/output tokens — substantially below GPT-5.5's $5/$30 and Opus 4.7's $5/$25, while delivering competitive coding-task quality.

Limitations

Not the Frontier on Hardest Tasks: Medium 3.5 is designed to compete on coding and agentic benchmarks at mid-tier price, not to lead frontier benchmarks like ARC-AGI-2 or GPQA Diamond against Gemini 3.1 Ultra, GPT-5.5 Pro, or Opus 4.7. For the hardest reasoning tasks, frontier closed models retain a meaningful edge.

License Complexity: Modified MIT introduces commercial-use terms beyond standard MIT — enterprises should review the specific terms before production deployment, particularly for services that resell or redistribute the model.

Mixed Independent Coverage: Some reviewers (e.g., revolutioninai.com) raised pricing concerns about Mistral Medium 3.5's $7.50/M output cost relative to its benchmark improvements over Medium 3 — Mistral's positioning is that the SWE-Bench result is the differentiator, but the value calculus depends on the workload.

Recent Developments

  • April 29, 2026 Launch: Released as open weights under Modified MIT license, with API availability simultaneously.
  • 77.6% on SWE-Bench Verified: The headline benchmark result, ahead of Devstral 2 and Qwen3.5 397B A17B.
  • Mistral 2026 Portfolio: Medium 3.5 sits alongside Mistral Small 4 (mistral-small-2603, hybrid instruct/reasoning/coding, 256K context), Mistral Large 3 (December 2025 flagship), Codestral, and Voxtral/Forge — completing Mistral's 2026 portfolio refresh.
  • Strategic Positioning: Medium 3.5 is Mistral's primary agentic and coding play, designed to compete in the open-weight tier alongside DeepSeek V4-Flash (released April 24, 2026), Qwen3.5 (released February 16, 2026), and Google's Gemma 4 (released April 2, 2026) — making April 2026 the most concentrated open-weight release window in recent memory.

Last Updated

May 7, 2026

→ Back to Models